ARRHYTHMIA CLASSIFICATION USING DEEP LEARNING METHODS IN THE CONTEXT OF HARDWARE LIMITATIONS OF WEARABLE DEVICES

Abstract

Long-term ECG monitoring with wearable devices offers broad opportunities for early, out-of-clinic detection of rhythm disorders, yet classification accuracy often remains below safety requirements. As populations age and arrhythmias become more prevalent, the need grows for reliable algorithms that can promptly identify high-risk conditions. The main obstacles in wearables are hardware constraints—few leads, lower sampling rates and bit depth, limited dynamic range and battery capacity—together with contact and motion artifacts. Under these conditions, a combined strategy is effective: deep-learning methods capture subtle morphological changes and long-range rhythmic dependencies, while orchestrated placement of computation between the device and the cloud reduces on-device load and extends battery life. The proposed approach aligns clinical relevance with the practical realities of wearables, specifies requirements for signal preparation, model architectures, and computational deployment, and provides a foundation for scalable services for arrhythmia screening and monitoring. The work also consolidates key ECG features and translates them into concrete requirements for data and models.